Whole-brain Perturb-seq atlas handed to distributed AI agents: thousands of parallel agents read it in minutes and recovered the source manuscript's conclusions
Synopsis
As biological experiments grow in scale, examining and interpreting the data becomes the rate-limiting step. Distributed biological reasoning assigns many AI agents to individual facets of an experiment and combines their findings into study-wide interpretations. Applied to a whole-brain Perturb-seq atlas, thousands of parallel agents examined the dataset in minutes against an estimated months of human analysis; without access to the source manuscript, the agents recovered its conclusions, predicted held-out perturbation effects, and identified perturbation relationships with shared or opposing outcomes across effect strengths.
Interpretation
The examination and interpretation of a biological experiment is split among many AI agents working in parallel, and their findings are combined into an interpretation covering the whole study rather than one agent reading all the data. The prior mode was single-agent operation, where literature coverage scaled only weakly with workload; in distributed mode the agents retrieved hundreds of times more references than single-agent mode. On a whole-brain Perturb-seq atlas, thousands of parallel agents examined the dataset in minutes, compared with an estimated months of human analysis.
The agents break the experiment into facets and reason over each separately, with each agent responsible for one part, and the facet-level conclusions are then pooled. This distributed biological reasoning is contrasted with single-agent mode, in which one agent carries out all reasoning and retrieval. Without access to the source manuscript, the agents recovered that manuscript's conclusions and predicted held-out perturbation effects.
Blinded human experts preferred the distributed-agent interpretations over single-agent interpretations and endorsed their discovery rankings. The preference is measured against interpretations produced by single-agent mode on the same task. The evaluation was blinded, and the experts also supported the distributed agents' discovery rankings.
Perspective
The demonstrated scope is one whole-brain Perturb-seq atlas: thousands of parallel agents examined it in minutes and, without the source manuscript, recovered its conclusions and predicted held-out perturbation effects; blinded human experts compared distributed and single-agent interpretations.
The human-analysis time is an estimate rather than a measurement; recovering the manuscript's conclusions shows agreement with them, not independent validation of their correctness; the blinded preference concerns interpretation quality and discovery rankings and is not reported as a quantitative score.
